MEDIA MANAGEMENT WEBINARS 2022

AI Facial Recognition

Face recognition at the Netherlands Institute for Sound and Vision - Is it really Harder, Better, Faster, Stronger: NISV implemented speaker detection and thesaurus label term extraction in 2015 in an effort to reduce time spent on manual annotation and to increase the amount of structured and fine grained access points for media professionals (the broadcasters) into our daily ingested programs. In 2019 face recognition has been thrown into the mix, trying to make automatic annotation “Harder, Better, Faster, Stronger”. Two years later it’s time to check the status: how well does it work, what’s the quality and quantity, what do our users think about it, what about the challenges on both technological and ethical aspects, and how does NISV apply it compared to other archives that might have similar challenges.

INA Facial recognition: Trombinos: We introduce Trombinos, a face database management system. It allows the user to navigate, cure and expand efficiently a large-scale dataset of faces, thanks to its integrated face recognition model. The database contains both face images coming from archived french TV shows, as well as face images obtained through image search engines. It already contains 62 million identified faces belonging to 70 thousand individuals, as well as 540 millions faces yet to be identified. Trombinos helps users to efficiently navigate this database, to recognize someone based on a picture, to retrieve their occurrences amongst the archived TV shows based on their name, and to manually enrich or correct the existing annotations. Our system facilitates human-machine interaction ; by identifying similar looking faces, Trombinos allows the users to validate them by batch and hence to annotate a large number of faces easily. Furthermore, Trombinos learns from the users feedback to improve its face recognition performances in real time, and in turn to facilitate the future annotations of the user in a virtuous circle. Using Trombinos, a single user can easily assign hundreds of thousands of faces to their correct identities in just a few minutes. While still in continuous development, Trombinos is already used regularly by 5 users at INA to annotate the database.

A look back at the use of facial recognition and other AI techniques at the RTS: AI was introduced and used at RTS for automatic metadata extraction a few years ago. This presentation is an opportunity to look back at the use of some of these tools, in particular facial recognition and visual search, and to discuss their impact in terms of process and changing practices.

AI User's Contributions Tools and Speech Regionalization Use Cases

Implementation of automation tools based on machine learning and artificial intelligence, what future does AI hold for audiovisual librarians, what role can they play and how to manage such a big change? In 2015, to deal with the challenges raised by the massification of digital data, INA has embarked on a vast project to completely overhaul its information systems.

Embracing a data centric strategy, this overhaul meets the necessary conditions for the implementation of automation tools based on machine learning and artificial intelligence to segment and describe audiovisual content. After identifying a few use cases, the works to automate segmentation and indexing media process was launch in 2019. These technical and material issues come with human and professional challenges, asking for specific business user’s skills and knowledges.

This presentation aims to give a few answers, at least advices and feedbacks, to a simple question: what future does AI hold for audiovisual librarians, what role can they play and how to manage such a big change?

Regional Languages with Machine Learning: Achievements and Challenges (small languages and dialects). Radiotelevisiun Svizra Rumantscha (RTR, SRG SSR) has launched a project for the written reproduction of spoken Romansh: The Romansh media company Radiotelevisiun Svizra Rumantscha (RTR, SRG SSR) has launched a project for the written reproduction of spoken Romansh. In order to be able to teach Romansh to the computer at all, RTR first had to acquire data to feed the system to learn Romansh. Bernard Bearth, Head of Data and Archives at RTR, explains the value of the existing archives for this project and the importance of speech to text and automatic translation for RTR archives in the future.

Regional Languages with Machine Learning: Achievements and Challenges University of Edinburgh : Gaelic Speech Recognition: Challenges Faced, Lessons Learnt and Future Plans: In this seminar, Dr Will Lamb (University of Edinburgh) will discuss work towards an accurate, general purpose Automatic Speech Recognition (ASR) system for Scottish Gaelic. In 2021, a team of researchers at Edinburgh developed the first working example of such as system, with assistance from MG Alba and several other non-academic partners. By the end of the project, the system achieved an accuracy level of approximately 73%. While this is a far cry from state-of-the-art English systems, used in applications such as Siri and Alexa, it is a promising beginning and confirms the methodology adopted. Modern ASR systems are machine-learning based, and trained upon tens of thousands, if not millions, of hours of audio and corresponding text. The principal challenge facing minority languages, such as Gaelic, is acquiring suitable training data. In this seminar, Dr Lamb will outline plans to augment the current 103 hours of data fiftyfold (i.e. to 5000 hours). This increase would translate to a much more accurate and robust system. The goal is to develop a subtitling system appropriate for broadcasting purposes in the next 3 to 4 years.

Metadata Gathering, Cloud Based Workflows and the Archivist - Where are we now?

Metadata wrangling at VRT wtih Tinkerlist: Managing metadata during TV production can be a very tedious and repetitive task — from manually uploading clips and adding numerous details (like names of the producers, creation dates, file types, etc.) to archiving them and hoping that one day you’ll be able to easily search and retrieve those clips.

Luckily, this no longer has to be the case. With TinkerList, production teams at broadcasting companies like VRT can automatically save and store metadata by simply uploading clips when they create scripts and rundowns, and entering the required details just once. The rest is archived and ready to retrieve in TinkerList’s database.

How exactly does it work and what else can TinkerList turn from ‘annoying’ to ‘automated’? Find out on the 26th of May during our talk with Erik Hauters and Aaron Nuytemans from TinkerList!

Cloud-based media supply chains: integration and evolution: The cloud has continued to rapidly evolve to support media workflows, including the distribution and monetization of content archives. This presentation will cover some recent updates to major cloud-based services relevant for media workflows since the last MMC Seminar in 2019, as well as some high-level strategy for turning archival focused media workflows into active media supply chains to support content reuse.

The Archvisit: The Impact of of AI /Machine Learning and the Pandemic on their Role: A wrap up and discussion focusing on if we can see a change of the archivist role in the context of AI /Machine Learning and the Pandemic. The discussion starts with an introduced from the perspectives that MMC has predicted and presented at the MMC Seminar in Stockholm 2019.